add reasoning content to ChatCompletions #494
Conversation
42aeffc to
5a88d0e
Compare
5a88d0e to
2ea9d85
Compare
| if not emit_reasoning_content: | ||
| emit_reasoning_content = True | ||
|
|
||
| reasoning_content_title = "# reasoning content\n\n" |
There was a problem hiding this comment.
this doesn't seem right - why hardcode?
There was a problem hiding this comment.
It's a markdown title for splitting the content and reasoning content. It's a constant value so have to hardcode.
The whole output are like below:
# reasoning content
Okay, I need to write a haiku about recursion in programming. Let's start by recalling what a haiku is. It's a three-line poem with a 5-7-5 syllable structure. The first line has five syllables, the second seven, and the third five again. The theme here is recursion in programming, so I should focus on elements that capture the essence of recursion.
.....
# content
**Haiku on Recursion:**
Function calls itself,
Base case breaks the looping chain—
Stack grows, then falls back.
Another way is use <think></think>
The whole output would be like this:
<think>
Okay, I need to write a haiku about recursion in programming. Let's start by recalling what a haiku is. It's a three-line poem with a 5-7-5 syllable structure. The first line has five syllables, the second seven, and the third five again. The theme here is recursion in programming, so I should focus on elements that capture the essence of recursion.
.....
</think>
**Haiku on Recursion:**
Function calls itself,
Base case breaks the looping chain—
Stack grows, then falls back.
Which way do you prefer?
There was a problem hiding this comment.
I think neither? IMO it would be better to emit a separate item for reasoning. For example, I was trying something like this in #581. What do you think?
There was a problem hiding this comment.
Yeah, I agree with you.
I just found out openai-python 1.7.6 already added the types needed for emit reasoning content.
from openai.types.responses import (
ResponseReasoningItem,
ResponseReasoningSummaryTextDeltaEvent,
ResponseReasoningSummaryPartAddedEvent,
ResponseReasoningSummaryPartDoneEvent,
ResponseReasoningSummaryTextDoneEvent
)
So we can emit ResponseReasoningSummaryTextDeltaEvent for reasoning content or create some class like ResponseReasoningTextDeltaEvent in this repo?
What do you think?
| if content is not None: | ||
| if not emit_content and is_reasoning_model: | ||
| emit_content = True | ||
| content_title = "\n\n# content\n\n" |
| @@ -62,9 +65,16 @@ async def handle_stream( | |||
| continue | |||
|
|
|||
| delta = chunk.choices[0].delta | |||
| reasoning_content = None | |||
| content = None | |||
| if hasattr(delta, "reasoning_content"): | |||
There was a problem hiding this comment.
when would this be true?
There was a problem hiding this comment.
When using reasoning model like deepseek-reasoner
deepseek reasoning model
OK |
fix issue #415